
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 7 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) (pow t_2 2.0)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ t_2 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_2, 2.0f);
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = t_2 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = sqrt(((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3)))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(t_2 / t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = floor(h) * dY_46_v; t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = sqrt(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = t_2 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + {t\_2}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 77.8%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
un-div-invN/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
Applied rewrites78.1%
Final simplification78.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) (pow t_2 2.0)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ 1.0 (/ t_4 t_2)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_2, 2.0f);
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = 1.0f / (t_4 / t_2);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = sqrt(((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3)))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(Float32(1.0) / Float32(t_4 / t_2)); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = floor(h) * dY_46_v; t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = sqrt(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = single(1.0) / (t_4 / t_2); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + {t\_2}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{t\_4}{t\_2}}\\
\end{array}
\end{array}
Initial program 77.8%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
un-div-invN/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
Applied rewrites78.1%
Final simplification78.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) (pow t_2 2.0)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (* t_2 (/ 1.0 t_4)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_2, 2.0f);
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = t_2 * (1.0f / t_4);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = sqrt(((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3)))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(t_2 * Float32(Float32(1.0) / t_4)); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = floor(h) * dY_46_v; t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = sqrt(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = t_2 * (single(1.0) / t_4); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + {t\_2}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{t\_4}\\
\end{array}
\end{array}
Initial program 77.8%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
un-div-invN/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
lift-/.f32N/A
/-rgt-identity78.0
Applied rewrites78.0%
Final simplification78.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_1 (* (floor h) dY.v))
(t_2 (+ (pow (* (floor w) dY.u) 2.0) (pow t_1 2.0)))
(t_3 (sqrt (fmax t_0 t_2))))
(if (>= t_0 t_2) (* (floor h) (/ dX.v t_3)) (* t_1 (/ 1.0 t_3)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_1 = floorf(h) * dY_46_v;
float t_2 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_1, 2.0f);
float t_3 = sqrtf(fmaxf(t_0, t_2));
float tmp;
if (t_0 >= t_2) {
tmp = floorf(h) * (dX_46_v / t_3);
} else {
tmp = t_1 * (1.0f / t_3);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_3 = sqrt(((t_0 != t_0) ? t_2 : ((t_2 != t_2) ? t_0 : max(t_0, t_2)))) tmp = Float32(0.0) if (t_0 >= t_2) tmp = Float32(floor(h) * Float32(dX_46_v / t_3)); else tmp = Float32(t_1 * Float32(Float32(1.0) / t_3)); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = ((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_1 = floor(h) * dY_46_v; t_2 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_1 ^ single(2.0)); t_3 = sqrt(max(t_0, t_2)); tmp = single(0.0); if (t_0 >= t_2) tmp = floor(h) * (dX_46_v / t_3); else tmp = t_1 * (single(1.0) / t_3); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + {t\_1}^{2}\\
t_3 := \sqrt{\mathsf{max}\left(t\_0, t\_2\right)}\\
\mathbf{if}\;t\_0 \geq t\_2:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{t\_3}\\
\end{array}
\end{array}
Initial program 77.8%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
un-div-invN/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
lift-*.f32N/A
pow2N/A
lift-pow.f3278.0
Applied rewrites78.0%
Applied rewrites77.8%
Final simplification77.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dX.v))
(t_4 (+ (pow t_0 2.0) (pow t_1 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (+ (* t_5 t_5) (* t_3 t_3)))
(t_7 (/ 1.0 (sqrt (fmax t_6 (+ (* t_0 t_0) t_2)))))
(t_8 (* t_3 t_7)))
(if (<= dX.u 0.10000000149011612)
(if (>= (pow t_3 2.0) t_4) t_8 (* t_1 t_7))
(if (>= (pow t_5 2.0) t_4)
t_8
(*
t_1
(/
1.0
(sqrt (fmax t_6 (+ t_2 (* dY.u (* dY.u (pow (floor w) 2.0))))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dX_46_v;
float t_4 = powf(t_0, 2.0f) + powf(t_1, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = (t_5 * t_5) + (t_3 * t_3);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, ((t_0 * t_0) + t_2)));
float t_8 = t_3 * t_7;
float tmp_1;
if (dX_46_u <= 0.10000000149011612f) {
float tmp_2;
if (powf(t_3, 2.0f) >= t_4) {
tmp_2 = t_8;
} else {
tmp_2 = t_1 * t_7;
}
tmp_1 = tmp_2;
} else if (powf(t_5, 2.0f) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = t_1 * (1.0f / sqrtf(fmaxf(t_6, (t_2 + (dY_46_u * (dY_46_u * powf(floorf(w), 2.0f)))))));
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(t_5 * t_5) + Float32(t_3 * t_3)) t_7 = Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(t_0 * t_0) + t_2) : ((Float32(Float32(t_0 * t_0) + t_2) != Float32(Float32(t_0 * t_0) + t_2)) ? t_6 : max(t_6, Float32(Float32(t_0 * t_0) + t_2)))))) t_8 = Float32(t_3 * t_7) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(0.10000000149011612)) tmp_2 = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = Float32(t_1 * t_7); end tmp_1 = tmp_2; elseif ((t_5 ^ Float32(2.0)) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(t_1 * Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(t_2 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))) : ((Float32(t_2 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))) != Float32(t_2 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0)))))) ? t_6 : max(t_6, Float32(t_2 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))))))))); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = floor(h) * dX_46_v; t_4 = (t_0 ^ single(2.0)) + (t_1 ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = (t_5 * t_5) + (t_3 * t_3); t_7 = single(1.0) / sqrt(max(t_6, ((t_0 * t_0) + t_2))); t_8 = t_3 * t_7; tmp_2 = single(0.0); if (dX_46_u <= single(0.10000000149011612)) tmp_3 = single(0.0); if ((t_3 ^ single(2.0)) >= t_4) tmp_3 = t_8; else tmp_3 = t_1 * t_7; end tmp_2 = tmp_3; elseif ((t_5 ^ single(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = t_1 * (single(1.0) / sqrt(max(t_6, (t_2 + (dY_46_u * (dY_46_u * (floor(w) ^ single(2.0)))))))); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := {t\_0}^{2} + {t\_1}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_5 \cdot t\_5 + t\_3 \cdot t\_3\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_0 \cdot t\_0 + t\_2\right)}}\\
t_8 := t\_3 \cdot t\_7\\
\mathbf{if}\;dX.u \leq 0.10000000149011612:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_3}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot t\_7\\
\end{array}\\
\mathbf{elif}\;{t\_5}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_2 + dY.u \cdot \left(dY.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)\right)}}\\
\end{array}
\end{array}
if dX.u < 0.100000001Initial program 80.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.0
Applied rewrites68.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3268.0
Applied rewrites68.0%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.0%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3268.9
Applied rewrites68.9%
if 0.100000001 < dX.u Initial program 70.5%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.0
Applied rewrites69.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3269.0
Applied rewrites69.0%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites69.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3269.1
Applied rewrites69.1%
Final simplification68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (+ (pow t_0 2.0) (pow t_1 2.0)))
(t_4 (* (floor w) dX.u))
(t_5
(/
1.0
(sqrt
(fmax (+ (* t_4 t_4) (* t_2 t_2)) (+ (* t_0 t_0) (* t_1 t_1))))))
(t_6 (* t_1 t_5))
(t_7 (* t_2 t_5)))
(if (<= dX.u 0.10000000149011612)
(if (>= (pow t_2 2.0) t_3) t_7 t_6)
(if (>= (pow t_4 2.0) t_3) t_7 t_6))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = powf(t_0, 2.0f) + powf(t_1, 2.0f);
float t_4 = floorf(w) * dX_46_u;
float t_5 = 1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1))));
float t_6 = t_1 * t_5;
float t_7 = t_2 * t_5;
float tmp_1;
if (dX_46_u <= 0.10000000149011612f) {
float tmp_2;
if (powf(t_2, 2.0f) >= t_3) {
tmp_2 = t_7;
} else {
tmp_2 = t_6;
}
tmp_1 = tmp_2;
} else if (powf(t_4, 2.0f) >= t_3) {
tmp_1 = t_7;
} else {
tmp_1 = t_6;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)) != Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2))) ? Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) : ((Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) != Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) ? Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)) : max(Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))))))) t_6 = Float32(t_1 * t_5) t_7 = Float32(t_2 * t_5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(0.10000000149011612)) tmp_2 = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= t_3) tmp_2 = t_7; else tmp_2 = t_6; end tmp_1 = tmp_2; elseif ((t_4 ^ Float32(2.0)) >= t_3) tmp_1 = t_7; else tmp_1 = t_6; end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = (t_0 ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = floor(w) * dX_46_u; t_5 = single(1.0) / sqrt(max(((t_4 * t_4) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1)))); t_6 = t_1 * t_5; t_7 = t_2 * t_5; tmp_2 = single(0.0); if (dX_46_u <= single(0.10000000149011612)) tmp_3 = single(0.0); if ((t_2 ^ single(2.0)) >= t_3) tmp_3 = t_7; else tmp_3 = t_6; end tmp_2 = tmp_3; elseif ((t_4 ^ single(2.0)) >= t_3) tmp_2 = t_7; else tmp_2 = t_6; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := {t\_0}^{2} + {t\_1}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_2 \cdot t\_2, t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}}\\
t_6 := t\_1 \cdot t\_5\\
t_7 := t\_2 \cdot t\_5\\
\mathbf{if}\;dX.u \leq 0.10000000149011612:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_2}^{2} \geq t\_3:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{elif}\;{t\_4}^{2} \geq t\_3:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.u < 0.100000001Initial program 80.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.0
Applied rewrites68.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3268.0
Applied rewrites68.0%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.0%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3268.9
Applied rewrites68.9%
if 0.100000001 < dX.u Initial program 70.5%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.0
Applied rewrites69.0%
lift-*.f32N/A
pow2N/A
lift-pow.f3269.0
Applied rewrites69.0%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites69.0%
Final simplification68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_2 t_2)) (+ (* t_0 t_0) (* t_1 t_1)))))))
(if (>= (pow t_2 2.0) (+ (pow t_0 2.0) (pow t_1 2.0)))
(* t_2 t_4)
(* t_1 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1))));
float tmp;
if (powf(t_2, 2.0f) >= (powf(t_0, 2.0f) + powf(t_1, 2.0f))) {
tmp = t_2 * t_4;
} else {
tmp = t_1 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)) != Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2))) ? Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) : ((Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) != Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) ? Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)) : max(Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(t_2 * t_4); else tmp = Float32(t_1 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1)))); tmp = single(0.0); if ((t_2 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = t_2 * t_4; else tmp = t_1 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_2 \cdot t\_2, t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_2}^{2} \geq {t\_0}^{2} + {t\_1}^{2}:\\
\;\;\;\;t\_2 \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot t\_4\\
\end{array}
\end{array}
Initial program 77.8%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.2%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3264.4
Applied rewrites64.4%
Final simplification64.4%
herbie shell --seed 2024232
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, v)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(if (>= (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor h) dX.v)) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor h) dY.v))))